User-Empowered Federated Learning in the Automotive Domain

This repo contains the code and experiments of the paper "User-Empowered Federated Learning in the Automotive Domain". Please, use this bib entry for referencing our work and derivatives:

@INPROCEEDINGS{10664305,
  author={Maugeri, Marcello and Paolo Morana, Mirko Ignazio and Esposito, Sergio and Bella, Giampaolo},
  booktitle={2024 IEEE International Conference on Blockchain (Blockchain)}, 
  title={User-Empowered Federated Learning in the Automotive Domain}, 
  year={2024},
  volume={},
  number={},
  pages={669-674},
  keywords={Training;Data privacy;Federated learning;Self-supervised learning;Flowering plants;Feature extraction;Data models;Federated Learning;Privacy;Automotive;Privacy-Enhancing Technologies},
  doi={10.1109/Blockchain62396.2024.00098}}

Disclaimer

The provided code is a proof of concept and is not intended for production use. The code is provided as is. Please, feel free to reach out to the authors for any questions or clarifications. Pull requests are welcome after the presentation of the paper.

Abstract

The proliferation of data generated through everyday device usage has prompted privacy concerns among users. In the automotive sector, this issue is particularly acute, given the substantial volumes of data collected in accordance with manufacturers' privacy policies. Privacy-Enhancing Technologies (PETs), such as Federated Learning (FL), offer a solution by safeguarding the confidentiality of car data while enabling decentralised machine learning model training, thus preventing the need for centralised data aggregation.

These FL-based models stand to benefit significantly from the diverse data distributions inherent in training across various features extracted from different cars. However, it remains imperative to ensure user awareness regarding their data processing, despite FL's privacy-preserving mechanisms.

To address this, we propose a User-Empowered FL approach, built upon the Flower Framework, empowering users to decide their participation in model training or merely inference without impacting the global model. We demonstrate this approach through an automotive case study utilising the EngineFaultDB dataset.

Finally, we outline future directions, particularly focusing on handling unlabelled data through self-supervised learning methodologies.

Install dependencies

pip install -r requirements.txt

Download the EngineFaultDB dataset

git clone https://github.com/Leo-Thomas/EngineFaultDB
cp EngineFaultDB/EngineFaultDB_Final.csv .
rm -rf EngineFaultDB

Prepare the dataset

cd utilities
cp ../EngineFaultDB_Final.csv .
python3 dataset_scaler.py
python3 dataset_splitter.py

The tflite model expects the features to be between 0 and 1. To achieve this, we use the MinMaxScaler from sklearn. After that, we split the dataset into two training sets and one test set (40/40/20). The train_1.csv, train_2.csv, and test_1.csv files are generated in the utilities folder. These must be copied to the client app under the path client/app/src/main/assets/data. You can use the shorthand command cp train_1.csv train_2.csv test_1.csv ../client/app/src/main/assets/data to copy the files.

Build the tflite model

cd utilities
python gen_tflite.py

Now you should have a enginefaultdb.tflite file in the gen_tflite folder. This must be copied to the client app under the path client/app/src/main/assets/model/enginefaultdb.tflite. You can use the shorthand command cp enginefaultdb.tflite ../client/app/src/main/assets/model/enginefaultdb.tflite to copy the file.

Run the server

python server.py

Run the client

The client is an Android Automotive app, please refer to the Android Automotive documentation to run the app on an emulator.

Reproduce the experiments

• Run the server
• Run three different clients, by setting the DEVICE_ID variable in the client app to 1, 2, and 3. This constant is found in the TabScreen.kt file.
• The clients will train the model for 10 local epochs, and the server will aggregate the models for 10 rounds.
• The main screen will show if an engine fault is detected in the car.

Expected results

We set the number of local epochs to 10, the number of round to 10 and batch size of 16. We managed to obtain an accuracy of ~75% on the test set.

Troubleshooting

• The generator for the tflite model requires tensorflow-cpu==2.9.2, if you have a different version of tensorflow installed, please create a virtual environment and install the required version. If your architecture does not support tensorflow-cpu (e.g. Apple Silicon), you can run the code on Colab.
• In MacOS there is a broken dependency in the flower_tflite code (protoc). To fix this, go in client/flower_tflite/build.gradle, comment line 36 and uncomment line 38.